Sicheng Liang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Sicheng Liang has a total of 33 co-authors. Here are the top co-authors by number of shared publications:
- Jie Wu
Fudan University, Shanghai, China
9 shared publications - Jingqi Feng
Fudan University, Shanghai, China
6 shared publications - Chunpu Huang
Fudan University, Shanghai, China
5 shared publications - Jiawei Ye
Fudan University
5 shared publications - Yukai Huang
Fudan University
5 shared publications - Ming Yan
Fudan University
4 shared publications
Latest Publications
- 18 August 2026
- Published by Elsevier in Knowledge-Based Systems
- Vol. 351
CoMAG: Collaborative GNN–LLM learning for multimodal attributed graphs
- 18 August 2026
- Published by Elsevier in Knowledge-Based Systems
- Vol. 351
CoMAG: Collaborative GNN–LLM learning for multimodal attributed graphs
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 1132-1142
OrionInfer: Low-Overhead Parallelism Switching and Live Migration for Efficient LLM Serving
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 1132-1142
Open AccessOrionInfer: Low-Overhead Parallelism Switching and Live Migration for Efficient LLM Serving
- Open Access
- 17 May 2026
- Published by MDPI AG in Future Internet
- Vol. 18 (5) , 265
Graph-Aware Scheduling for Multi-Agent Workflows in Edge-Cloud Environments
Multi-agent workflows have emerged as an important execution pattern for intelligent applications, where specialized agents collaborate through dependent stages such as planning, retrieval, execution, and verification. When such workflows are deployed over edge-cloud infrastructures, scheduling becomes challenging because task dependencies, heterogeneous resource conditions, and topology-dependent communication overhead must be considered jointly. We study the problem of scheduling multi-agent workflows in edge-cloud environments and propose a graph-aware scheduling method that models workflow execution as a task graph and the underlying infrastructure as a resource graph. The method combines structure-aware task and resource representations, communication-sensitive assignment scoring, and online resource-state updates to improve task placement quality. Experiments across different workflow complexities, system scales, and dynamic operating conditions show that the proposed method achieves a favorable balance among latency, communication overhead, and execution cost, particularly in communication-sensitive and medium-to-large-scale settings. These results suggest that jointly modeling workflow structure and resource topology can improve scheduling quality for multi-agent workflow execution in heterogeneous edge-cloud environments.
- 17 May 2026
- Published by MDPI AG in Future Internet
- Vol. 18 (5) , 265
Open AccessGraph-Aware Scheduling for Multi-Agent Workflows in Edge-Cloud Environments
Multi-agent workflows have emerged as an important execution pattern for intelligent applications, where specialized agents collaborate through dependent stages such as planning, retrieval, execution, and verification. When such workflows are deployed over edge-cloud infrastructures, scheduling becomes challenging because task dependencies, heterogeneous resource conditions, and topology-dependent communication overhead must be considered jointly. We study the problem of scheduling multi-agent workflows in edge-cloud environments and propose a graph-aware scheduling method that models workflow execution as a task graph and the underlying infrastructure as a resource graph. The method combines structure-aware task and resource representations, communication-sensitive assignment scoring, and online resource-state updates to improve task placement quality. Experiments across different workflow complexities, system scales, and dynamic operating conditions show that the proposed method achieves a favorable balance among latency, communication overhead, and execution cost, particularly in communication-sensitive and medium-to-large-scale settings. These results suggest that jointly modeling workflow structure and resource topology can improve scheduling quality for multi-agent workflow execution in heterogeneous edge-cloud environments.
- Open Access
- 8 May 2026
- Published by Elsevier in SSRN Electronic Journal
CoMAG: Collaborative GNN–LLM Learning for Multimodal Attributed Graphs
Multimodal Attributed Graphs (MAGs), which include textual, visual, and structural information, are prevalent in domains such as e-commerce and social media. However, zero-shot reasoning on MAGs remains challenging. Existing graph neural networks often produce biased representations due to modality imbalance and limited semantic reasoning, and their reliance on labeled data makes them almost incapable of handling such scenarios. Meanwhile, large language models (LLMs) exhibit strong cross-modal reasoning ability but lack explicit modeling of graph structure. To address these challenges, we propose CoMAG, a collaborative framework that unifies graph-enhanced multimodal representation learning with LLM-based reasoning. CoMAG extracts modality-specific embeddings via frozen encoders and refines them through self-supervised graph learning. It further derives discrete structural priors as symbolic tokens and aligns graph embeddings with the LLM token space, enabling prompt-based inference. Moreover, a structure-aware reasoning layer introduces gated structural cues, adaptive modality fusion, and neighborhood-consistent prediction to enhance robustness. Extensive experiments on multiple real-world MAG datasets for zero-shot tasks demonstrate the effectiveness of CoMAG, yielding consistent and notable improvements over state-of-the-art baselines. These results highlight CoMAG’s capability to tackle the fundamental challenges of zero-shot multimodal graph reasoning.
- 8 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessCoMAG: Collaborative GNN–LLM Learning for Multimodal Attributed Graphs
Multimodal Attributed Graphs (MAGs), which include textual, visual, and structural information, are prevalent in domains such as e-commerce and social media. However, zero-shot reasoning on MAGs remains challenging. Existing graph neural networks often produce biased representations due to modality imbalance and limited semantic reasoning, and their reliance on labeled data makes them almost incapable of handling such scenarios. Meanwhile, large language models (LLMs) exhibit strong cross-modal reasoning ability but lack explicit modeling of graph structure. To address these challenges, we propose CoMAG, a collaborative framework that unifies graph-enhanced multimodal representation learning with LLM-based reasoning. CoMAG extracts modality-specific embeddings via frozen encoders and refines them through self-supervised graph learning. It further derives discrete structural priors as symbolic tokens and aligns graph embeddings with the LLM token space, enabling prompt-based inference. Moreover, a structure-aware reasoning layer introduces gated structural cues, adaptive modality fusion, and neighborhood-consistent prediction to enhance robustness. Extensive experiments on multiple real-world MAG datasets for zero-shot tasks demonstrate the effectiveness of CoMAG, yielding consistent and notable improvements over state-of-the-art baselines. These results highlight CoMAG’s capability to tackle the fundamental challenges of zero-shot multimodal graph reasoning.
- Open Access
- 26 January 2026
- Published by Springer Nature in Journal of King Saud University Computer and Information Sciences
- Vol. 38 (3)
LACE: Mitigating cold-starts in serverless with a multi-task mixture-of-experts caching
- 26 January 2026
- Published by Springer Nature in Journal of King Saud University Computer and Information Sciences
- Vol. 38 (3)
Open AccessLACE: Mitigating cold-starts in serverless with a multi-task mixture-of-experts caching
- Open Access
StyleBreak: Revealing Alignment Vulnerabilities in Large Audio-Language Models via Style-Aware Audio Jailbreak
Large Audio-language Models (LAMs) have recently enabled powerful speech-based interactions by coupling audio encoders with Large Language Models (LLMs). However, the security of LAMs under adversarial attacks remains underexplored, especially through audio jailbreaks that craft malicious audio prompts to bypass alignment. Existing efforts primarily rely on converting text-based attacks into speech or applying shallow signal-level perturbations, overlooking the impact of human speech's expressive variations on LAM alignment robustness. To address this gap, we propose StyleBreak, a novel style-aware audio jailbreak framework that systematically investigates how diverse human speech attributes affect LAM alignment robustness. Specifically, StyleBreak employs a two-stage style-aware transformation pipeline that perturbs both textual content and audio to control linguistic, paralinguistic, and extralinguistic attributes. Furthermore, we develop a query-adaptive policy network that automatically searches for adversarial styles to enhance the efficiency of LAM jailbreak exploration. Extensive evaluations demonstrate that LAMs exhibit critical vulnerabilities when exposed to diverse human speech attributes. Moreover, StyleBreak achieves substantial improvements in attack effectiveness and efficiency across multiple attack paradigms, highlighting the urgent need for more robust alignment in LAMs.
Open AccessStyleBreak: Revealing Alignment Vulnerabilities in Large Audio-Language Models via Style-Aware Audio Jailbreak
Large Audio-language Models (LAMs) have recently enabled powerful speech-based interactions by coupling audio encoders with Large Language Models (LLMs). However, the security of LAMs under adversarial attacks remains underexplored, especially through audio jailbreaks that craft malicious audio prompts to bypass alignment. Existing efforts primarily rely on converting text-based attacks into speech or applying shallow signal-level perturbations, overlooking the impact of human speech's expressive variations on LAM alignment robustness. To address this gap, we propose StyleBreak, a novel style-aware audio jailbreak framework that systematically investigates how diverse human speech attributes affect LAM alignment robustness. Specifically, StyleBreak employs a two-stage style-aware transformation pipeline that perturbs both textual content and audio to control linguistic, paralinguistic, and extralinguistic attributes. Furthermore, we develop a query-adaptive policy network that automatically searches for adversarial styles to enhance the efficiency of LAM jailbreak exploration. Extensive evaluations demonstrate that LAMs exhibit critical vulnerabilities when exposed to diverse human speech attributes. Moreover, StyleBreak achieves substantial improvements in attack effectiveness and efficiency across multiple attack paradigms, highlighting the urgent need for more robust alignment in LAMs.
- 4 October 2025
- Published by Springer Nature
- p. 176-192
HCT: A Hierarchical Contrastive Learning Framework for Transferable Graph Anomaly Detection
- 4 October 2025
- Published by Springer Nature
- p. 176-192
HCT: A Hierarchical Contrastive Learning Framework for Transferable Graph Anomaly Detection
- Open Access
- 3 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 1707-1717
Feature Reconstruction for Anomaly Detection on Directed Multigraphs: A Preprocessing Framework for GNNs
- 3 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 1707-1717
Open AccessFeature Reconstruction for Anomaly Detection on Directed Multigraphs: A Preprocessing Framework for GNNs
- Open Access
- 20 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 1283-1295
WindServe: Efficient Phase-Disaggregated LLM Serving with Stream-based Dynamic Scheduling
- 20 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 1283-1295
Open AccessWindServe: Efficient Phase-Disaggregated LLM Serving with Stream-based Dynamic Scheduling
- Open Access
- 12 August 2024
- Published by Association for Computing Machinery (ACM)
- p. 929-938
Revisiting Learned Index with Byte-addressable Persistent Storage
Byte-addressable Persistent Storage (BPS), such as persistent memory and CXL-enabled SSDs, has become an extension of main memory. This opens up new possibilities for indexes that operate and persist data directly on the memory bus. Recent learned indexes exploit data distribution and have shown great potential for some workloads. Despite some work proposed for integrating learned indexes into BPS, they are mainly based on Intel’s first-generation persistent memory. The current design suffers from the following problems: 1) Excessive storage line accesses due to large node in learned indexes; 2) Inefficient concurrency control due to volatile cache; 3) Write amplification due to mismatch access granularity. To resolve these challenges, we introduce a new learned index named PFLX, featuring three key design improvements: 1) a Storage Line-Friendly Node Layout that enhances search and insertion operations and minimizes storage line accesses through strategic use of pointers; 2) a Persistent Cache-based Lock-free Concurrency Mechanism that utilizes atomic primitives to serialize operations in leaf nodes effectively; 3) a Selective Data Flush Mechanism designed to reduce write amplification. Evaluations show that PFLX outperforms the state-of-the-art persistent indexes by 1.2 ∼ 3 ×.
- 12 August 2024
- Published by Association for Computing Machinery (ACM)
- p. 929-938
Open AccessRevisiting Learned Index with Byte-addressable Persistent Storage
Byte-addressable Persistent Storage (BPS), such as persistent memory and CXL-enabled SSDs, has become an extension of main memory. This opens up new possibilities for indexes that operate and persist data directly on the memory bus. Recent learned indexes exploit data distribution and have shown great potential for some workloads. Despite some work proposed for integrating learned indexes into BPS, they are mainly based on Intel’s first-generation persistent memory. The current design suffers from the following problems: 1) Excessive storage line accesses due to large node in learned indexes; 2) Inefficient concurrency control due to volatile cache; 3) Write amplification due to mismatch access granularity. To resolve these challenges, we introduce a new learned index named PFLX, featuring three key design improvements: 1) a Storage Line-Friendly Node Layout that enhances search and insertion operations and minimizes storage line accesses through strategic use of pointers; 2) a Persistent Cache-based Lock-free Concurrency Mechanism that utilizes atomic primitives to serialize operations in leaf nodes effectively; 3) a Selective Data Flush Mechanism designed to reduce write amplification. Evaluations show that PFLX outperforms the state-of-the-art persistent indexes by 1.2 ∼ 3 ×.